Group comparisons in logit and probit

using predicted probabilities1

J. Scott Long
Indiana University
June 25, 2009

Abstract
The comparison of groups in regression models for binary outcomes is
complicated by an identification problem inherent in these models. Tra-
ditional tests of the equality of coefficients across groups confound the
magnitude of the regression coefficients with residual variation. If the
amount of residual variation differs between groups, the test can lead
to incorrect conclusions (Allison 1999). Allison proposes a test for the
equality of regression coefficients that removes the effect of group differ-
ences in residual variation by adding the assumption that the regression
coefficients for some variables are identical across groups. In practice, a
researcher is unlikely to have either empirical or theoretical justification
for this assumption, in which case the Allison’s test can also lead to incor-
rect conclusions. An alternative approach, suggested here, uses predicted
probabilities. Since predicted probabilities are unaffected by residual vari-
ation, tests of the equality of predicted probabilities across groups can be
used for group comparisons without assuming the equality of the regres-
sion coefficients of some variables. Using predicted probabilities requires
researchers to think differently about comparing groups. With tests of the
equality of regression coefficients, a single test lets the researcher conclude
easily whether the effects of a variable are equal across groups. Testing the
equality of predicted probabilities requires multiple tests since group dif-
ferences in predictions vary with the levels of the variables in the model.
A researcher must examine group differences in predictions at multiple
levels of the variables often requiring more complex conclusions on how
groups differ in the effect of a variable.

1
I thank Paul Allison, Ken Bollen, Rafe Stolzenberg, Pravin Trivedi, and Rich Williams for their
comments.

1
 Group comparisons in logit and probit
using predicted probabilities

1 Overview
The comparison of groups is fundamental to research in many areas and tests com-
paring groups have received a great deal of attention. Chow’s (1960) paper, declared
a “citation classic” (Garfield 1984), provides a general framework for group com-
parisons in the linear regression model. Suppose that we are comparing the effect
of  on  for women and men, where   
 and   are the coefficients of interest.
 
If 0 :   =   is rejected, we conclude that the effect of  differs for men and
women. This approach to testing group differences can be applied to many types
of regression models as shown by Liao (2002). Allison (1999) points out a critical
problem when this test is used in models such as logit or probit. For these models,
standard tests can lead to incorrect conclusions since they confound the magnitude
of the regression coefficients with the amount of residual variation. Allison proposes
a test that removes the effect of residual variation by assuming that the coefficients
for at least one independent variable are the same in both groups. Unfortunately, a
researcher might lack sufficient theoretical or empirical information to justify such an
assumption. Making an ad hoc decision that some regression coefficients are equal
can lead to incorrect conclusions.2 Tests of predicted probabilities provide an al-
ternative approach for comparing groups that is unaffected by group differences in
residual variation and does not require assumptions about the equality of regression
coefficients for some variables. Group comparisons are made by testing the equality
of predicted probabilities at different values of the independent variables.
This paper begins by reviewing why standard tests of the equality of regression
coefficients across groups are inappropriate in some types of models. I then show
why predicted probabilities are unaffected by residual variation and present a test
of the equality of predicted probabilities that can be used for group comparisons of
effects in models such as logit and probit. To illustrate this approach, I begin with
a model that includes a single independent variable, in which case all information
about group differences can be shown in a simple graph. I extend this approach to
models with multiple independent variables, which requires more complex analysis
due to the nonlinearity of the models.
2
Williams (2009) raises other concerns with this tests that are discussed below.

2
2 Model identification in logit and probit
Discussions of the logit and probit model often note that the slope coefficients are
only identified up to a scale factor (Maddala 1983:23). This lack of identification is
why standard tests of the equality of regression coefficients across groups should not
be used. To see how identification causes a problem in group comparisons and why
predicted probabilities are not affected by this problem, consider how these models are
derived using an underlying latent variable (see Long 1997: 40-50 for a full derivation).
Suppose that the latent  ∗ is linearly related to an observed  through the structural
model
 ∗ = 0 + 1  + 
where I use a single independent variable for simplicity. The latent  ∗ is linked to an
observed, binary  by the measurement equation
½
1 if  ∗  0
= (1)
0 if  ∗ ≤ 0 
If  ∗ is greater than 0,  is observed as 1. Otherwise,  is observed as 0. For example,
when a person’s propensity to be in the labor force exceeds 0, she joins the labor
force (i.e.,  = 1). If her propensity is at or below 0, she is not in the labor force (i.e.,
 = 0).
According to equation 1, the probability of  = 1 is the proportion of the distrib-
ution of  ∗ that is above 0 at a given value of :

Pr ( = 1 | ) = Pr ( ∗  0 | )  (2)

Substituting  ∗ = 0 + 1  +  and rearranging terms, the probability can be written

in terms of the distribution of the errors:

Pr( = 1 | ) = Pr( ≤ 0 + 1  | )

[ Figures 1 & 2 about here ]

Figure 1 illustrates the relationship between the structural model for  ∗ and the
probability of  for specific values of the parameters labeled as set :

∗ =  
0 + 1  +  (3)
= −6 + 1 +  

Assume that  is normally distributed with mean 0 and variance  2 = 1, which are
the usual assumptions for the probit model. The probability of  = 1 at a given value
of  is the shaded portion of the error distribution in Figure 1 corresponding to

Pr ( = 1 | ) = Pr( ≤  
0 + 1  | ) (4)

3
If ∗ was observed, as is the case for the dependent varaible in the linear regres-
sion model, we could estimate the variance of  . But, since both  and ∗ are
unobserved, the variance of  is unidentified and must be assumed. The lack of
identification can be shown by changing the parameters in equation 3 by any non-
zero factor :
∗ =  
0 + 1  +  .

For example, suppose that  = 2. Then

2∗ = 2 
0 + 21  + 2 .

Defining ∗ ≡ 2∗ ,    

0 ≡ 20 , 1 ≡ 21 , and  ≡ 2 , the structural equation can
be written as

∗ =  
0 + 1  +  (5)
= −12 + 2 + 

where   = 2  = 2. While the slope, intercept, and standard deviation of the

errors in equations 3 and 5 differ by a factor of  = 2, the predicted probabilities are
unaffected since changing the parameters by a scale factor simply “stretches” Figure
1, as shown in Figure 2. With set  of parameters, we are multiplying the inequality
in equation 4 by  = 2, which does not change the probability of  = 1:

Pr ( = 1 | ) = Pr(2 ≤ 2 
0 + 21  | )
= Pr( ≤  
0 + 1  | ) 

While the magnitude of the ’s and  are affected by the factor change, the predicted
probabilities are not. Even if the magnitude of the ’s and  are of substantive
interest, they cannot be estimated individually since all scale factors of the parameters
have the same predictions for the observed data. That is, the ’s are only identified
up to a scale factor. Typically, to identify the model, the variances of the errors are
assumed to have a single, specific value.
This is the fundamental issue behind Allison’s critique of standard tests to compare
coefficients across groups in logit and probit models. To see this, consider a pair of
models for men and women:
¡ ¢
Men:  ∗ =  
0 + 1  +  where  ∼  0  
2
(6)
∗  
¡ 2
¢
Women:  = 0 + 1  +  where  ∼  0   (7)

Suppose that the slopes for  are identical, but that residual variation is greater for
women than men. That is,   2 2 ∗
1 = 1 and      . If  was observed, this would be
a pair of linear regressions with identical slopes for . Since the variance of the errors
is larger for women than men, the coefficient of determination 2 would be smaller

4
for women. But, since  ∗ is latent, the variance of errors must be assumed. While
the variance could be fixed to any value, the usual (but arbitrary) assumption for
probit is that  2 = 1 and for logit that  2 =  2 3. Here, for simplicity and without
loss of generality, I assume that  2 = 1. Critically, the same identifying assumption
is used for both men and women. That is, both men and women are assumed to have
the save variance for  and accordingly the same 2 . To apply this assumption to
equation 6 (i.e., to force the variance of  to be 1), I divide the equation by   . For
men,
∗   
= 0 + 1 + .
   
To simplify notation, I use ’s for the rescaled parameters and  for the rescaled error:
∗
=  
0 + 1  +   (8)

where the variance of   is 1. Similarly for women, I transform equation 7:

∗ 
0  
= + 1 +
   
 
= 0 + 1  +   (9)

where   also has a unit variance by construction.

Software for probit as found in standard packages such as Stata, SPSS and SAS
incorporates the identifying assumption that the variance of the errors is 1 and esti-
mates equations 8 and 9, not equations 6 and 7.3 That is, the software estimates the
rescaled -parameters, not the original -parameters. A standard test of the equality
of regression coefficients across groups evaluates the hypothesis


1 
 : = 1
 
or equivalently
 :   
1 = 1 ,

not the hypothesis

 :  
1 = 1 .

Moreover,  and  are only equivalent if   =   . This is why standard tests

confound the slopes and the variances of the errors which can lead to incorrect con-
clusions.
To test the cross group equality of the effect of  without confounding group dif-
ferences in residual variation, that is, to test  :  
1 = 1 , requires information
3
Standard software for logit rescales the errors to  = √ so that Var( ) =  2 3
3

5
about group differences in the variance of the errors. Allison’s test obtains this in-
formation by assuming that the coefficients for at least one variable are the same for
both groups. For example, suppose that our model also include the predictor :
¡ ¢
Men:  ∗ =   
0 + 1  + 2  +  where  ∼  0  
2
¡ ¢
Women:  ∗ =   
0 + 1  + 2  +  where  ∼  0   
2

If the slope coefficient for  is the same for both groups (i.e.,  
2 = 2 = 2 ), then
the ratio of the rescaled -coefficients for  equals the ratio of the standard deviations
of the errors: ¡  ¢

2  2   (2   ) 

=  = = .
2 (2   ) (2   ) 
This provides the information needed to test  :  
1 = 1 .

[ Figure 3 about here ]

There are two issues to keep in mind when using Allison’s test. First, the test
depends critically on the assumption that the effects of at least one variable are equal
across groups. If the equality assumption is justified, Allison’s test allows you to test
the equality of the -coefficients across groups in logit and probit without contam-
ination by group differences in residual variation. In practice, however, I think it
is unlikely that researchers will have sufficient information to justify the assumption
that the effects of some variables are zero. If the equality assumption is unjusti-
fied, Allison’s test can lead to incorrect conclusions. Using the same basic approach,
Williams (2009) extends and refines the tests proposed by Allison by showing how
they can be incorporated into the heterogeneous choice model.
A second issue related to Allison’s approach is that the equality of regression
coefficients across groups has different implications in nonlinear models than in linear
models. In a linear model, if the coefficient for  is the same for men and women, both
men and women receive the same expected change in  for a given change in  holding
all other variables constant. For example, for each additional year of education both
men and women are expected to increase their wages by   dollars, holding all else
constant.4 In a nonlinear model, having the same   for two groups does not imply
that a change in  will have the same change in Pr ( = 1) for both groups. This is
shown in Figure 3 which plots a probit or logit model where men and women have
the same coefficient for  while the intercepts differ. As  increases from 2 to 3,
the change in predicted probability for men, ∆Men , is much larger than the change
∆Women for women. In nonlinear models, having equal regression parameters for a
given variable does not imply similar changes in the predicted probabilities.
4
Significantly, in the linear regression model the equality of regression coefficients for two groups
does not imply that the 2 ’s for two groups are the same.

6
3 Testing group differences in predicted probabil-
ities
While regression coefficients are affected by the identifying assumption for the vari-
ance of the errors, the predicted probabilities are not. Accordingly, if you estimate
logit and probit
√ models using the same data, the estimated ’s for the logit model will
be about  3 times larger than those from the√probit model because the assumed
standard deviation of the errors for logit is  3 times larger than the assumed
standard deviation for probit. The predicted probabilities, however, will be nearly
identical. They differ only because the shape of the normal distribution used in probit
is slightly different than the shape of the logistic distribution assumed for logit. Since
predicted probabilities are not affected by group differences in residual variation, you
can compare groups by testing the equality of predicted probabilities at substantively
interesting values of the independent variables.
Tests of predicted probabilities can be computed with the delta method (Xu and
Long 2005). The delta method computes the variance of functions of maximum
likelihood estimates by creating a linear approximation of a function and computing
the variance of the simpler, linear function (Cameron and Trivedi 2005:227-233). Let
b be a vector of maximum likelihood estimates and let  () be a function of β, such
β
as predicted probabilities. Using a Taylor series expansion, (β) b ≈ (β) + () (βb−
0
β). Under standard assumptions, (β) b is distributed normally around (β) with a
variance of
h i b b
Var (β) b = (β)  (β) b (β)  (10)
βb0 b
β
where  (β)b is the covariance matrix for the estimated parameters.
To apply the delta method to predicted probabilities, it helps to use the shorthand
notation that  (x ) ≡ Pr ( = 1 | x ). If  (β) =  (x ) =  (x0 β) where  () is the
cdf for the standard normal distribution for the probit model and the standardized
logistic distribution for the logit model. Then

(β)  (x0 β)
=
β β
=  (x0 β) x

where  () is the pdf. Then,

 (x0 β) b  (x0 β)
Var [ (x )] =  (β)
β0 β
0 0 b
=  (x β) x  (β)x  (x0 β) 

7
The variance of the group differences in probabilities is
h i h i h i
∗ Group 1 ∗ Group 2 ∗ Group 1 ∗ Group 2
Var  (x1 ) −  (x2 ) = Var  (x1 ) + Var  (x2 ) 

where x∗1 and x∗2 indicate values of the ’s for the two groups. These values could be
identical (e.g., test the predicted probabilities for the same levels of all variables) or
different (e.g., test the predicted probabilities when the values of the ’s for group 1
are the means for group 2 and vice versa). The -statistic to test 0 :  (x∗1 )Group 1 =
 (x∗2 )Group 2 is
 (x∗1 )Group 1 −  (x∗2 )Group 2
=r h i , (11)
∗ Group 1 ∗ Group 2
Var  (x1 ) −  (x2 )

which has an asymptotic normal distribution. Defining ∆ (x) ≡  (x∗1 )Group 1 −

 (x∗2 )Group 2 , a 95% confidence interval can be constructed such that
³ ´
Pr ∆ (x)Lower Bound ≤ ∆ (x) ≤ ∆ (x)Upper Bound = 95 

Further details on the derivation are provided in the appendix.

4 Example
The following example is based on 301 male and 177 female biochemists who obtained
Ph.D.s in the late 1950s and early 1960s and who held faculty positions in graduate
departments (see Long, Allison, and McGinnis (1993) for further details). Following
Allison (1999), the unit of analysis is person-years with each observation correspond-
ing to one year in rank for one scientist with multiple records for each scientist (one
record for each year a scientist is an assistant professor). The dependent variable is
coded 1 if a scientist obtained tenure in the given year, else 0. Independent variables
include the number of years a scientist has been in the rank of assistant professor,
the selectivity of the scientist’s undergraduate institution, the cumulative number of
articles published by the end of each person-year, and the prestige of the department
in which the scientist is working. Descriptive statistics are given in Table 1. Using
these data, three increasingly complex models are estimated and groups differences
examined. Model 1 uses only gender and the number of articles to predict tenure.
Model 2 adds a binary indictor of having a job in a distinguished department. Model
3 includes year, year-squared, selectivity, total articles, and departmental prestige.
Analyses were completed using Stata Version 10 (2007) along with Long and Freese’s
(2005) SPost commands.5
5
The GC (group comparison) package by Long(2009) is a wrapper for the SPost command
prvalue that makes estimating group comparisons much simpler.

8
 [ Tables 1 & 7 and Figures 4 & 5 about here ]

Model 1 uses only the number of articles to predict tenure:

  (articles) = Λ(  
0 +  1 articles)
 (articles) = Λ(  
0 +  1 articles)

with estimated ’s given in Table 2. Based on these estimates, the predicted prob-
ability of tenure for men with 15 articles is .23 and for women is .14, a difference of
.09. To test if the difference is significant, I compute a -test using equation 11:

b (articles=15 ) − 
 b (articles=15 )
 = q £  ¤
b (articles=15 ) − 
Var  b (articles=15 )
= 411

I reject the null hypothesis 0 :  (articles=15 ) =   (articles=15 ) at the .01 level.

Since the logit model is nonlinear, gender differences in the predicted probability
of tenure vary by the number of articles. This is shown in Figure 4 which plots
the probability of tenure for men and women with articles ranging from 0 to 50.
Differences in probability are small for scientists with fewer publications. For example,
for scientists with 7 publications, the difference is only .016 which is not significantly
different from 0 ( = 132). To examine gender differences throughout the range of
b (articles)−b
articles, I can plot   (articles) along with the 95% confidence interval
as shown in Figure 5. If the confidence interval extends below 0, the difference
is not significant. Overall, gender differences in the probability of tenure are not
significant for those with less than 10 publications and get steadily larger until about
40 publications at which point they level off with a difference of about .45.
This simple example illustrates a critical difference between tests of regression co-
efficients and tests using predicted probabilities. With tests of regression coefficients,
there is a single test of whether the effect of a variable differs across groups.6 With
predicted probabilities, there are multiple tests for different levels of the independent
variables. As we just saw, the group differences can be significant at some values of
the predictor, while not significant at others. This complicates things for the data
analyst since there might not be a simple answer to the question of whether the effects
of a variable are the same for both groups. But there are also advantages. First, using
probabilities avoids assumptions about the equality of the effects of other variables.
Second, the more complex answer to whether the effects of a variable differ across
groups could be a more realistic representation of the phenomenon being modeled. In
6
Note that Allison’s test can not be used with a single predictor since there is no other variable
in the model whose effect can be assumed equal across groups.

9
this example, it makes substantive sense that gender differences are small and non-
significant at lower levels of productivity since neither men nor women are likely to
be tenured if they have not been productive, but there could gender discrimination
at higher levels of productivity where tenure is more likely.

[ Table 3 and Figures 6, 7 & 8 about here ]

With a single independent variable, a simple graph shows all of the differences
between men and women in the effect of articles on tenure. With additional vari-
ables, interpretation is more complicated since group differences in predictions differ
by the level of all variables in the model. This is a critical point for understanding
how probabilities are used to compare groups. To show this, Model 2 retains the
number of articles and adds a binary variable indicating whether a scientist is lo-
cated in a distinguished department, leading to the estimates in Table 3. While my
substantive interest may be in how productivity affects tenure for men and women,
gender differences in the probability of tenure for a given number of articles depends
on whether a faculty member works in a distinguished department. To show this,
Figure 6 plots gender differences in the probability of tenure for those not in dis-
tinguished departments while Figure 7 plots differences for those in distinguished
departments. These figures show that gender differences in the probability of tenure
are noticeably larger in distinguished departments for faculty with more than twenty-
five articles. To combine the two figures and to introduce a technique that is essential
for more complex models, I use a bold line to plot gender differences for those from
distinguished departments and a thin line for those who are not. If the difference
in probability is significant, the line is solid; if the difference is not significant, the
line is dashed. Figure 8 shows how the prior graphs can be combined. With two
predictors, not only isn’t there a single test of the effect of articles, but now the tests
for differences in probabilities depend on the level of articles and the prestige of the
department. This complexity, however, may also provide interesting insights into the
tenure process, suggesting that gender differences in the tenure process are greater in
more prestigious departments.

[ Table 4 and Figures 9, 10 & 11 about here ]

As a more realistic example, I estimate the full model used in Allison (1999)
with estimates shown in Table 4. Gender differences in the probability of tenure
differ by the levels of five variables. Even with only five variables, which is a small
number compared to the number of variables in many applications, it is impractical
to examine all combinations of the variables. Instead, the examine predictions you
must focus on those regions of the data space that are substantively most interesting.
For example, in most universities tenure decisions are made in the seventh year in
rank, so I compute predicted probabilities holding year at 7 and year-squared at 49.

10
As with Model 2, I am interested in the effects of articles and prestige. Since prestige
is now measured as a continuous variable ranging from 1 to 5, I extend Figure 8 for
Model 2 to show gender differences at five levels of prestige over a range of articles
from 0 to 50. While I want to control for the selectivity of the baccalaureate, I am not
particularly interested in its effect, so I hold this variables at the mean. Predictions
are made by varying the number of articles and the level of prestige in these equations:

  (articles, prestige) = Λ(    

0 +  1 (7) +  2 (49) +  3 selectivity
+  
4 articles +  5 prestige)
  (articles, prestige) = Λ(    
0 +  1 (7) +  2 (49) +  3 selectivity
+  
4 articles +  5 prestige)

Varying articles for five levels of prestige leads to Figure 9.7 Gender differences are
smaller at lower levels of productivity, but the less prestigious the department the
greater the difference. Since less prestigious departments are likely to require fewer
publications for tenure than more prestigious departments, this is what would be
expected. At higher levels of productivity, the differences are larger, with larger
differences at more prestigious departments. This figure contains information on how
both productivity and prestige affect tenure and our interpretation reflects the effects
of both variables. To better understand how the effects differ by both prestige and
productivity, I can plot the same predictions letting prestige vary continuously from
1 to 5 at fixed levels of productivity. This is shown Figure 10. In my experience, most
people who first look at Figure 9 are surprised by Figure 10. When examining graphs
of differences in predicted probabilities it is important to keep in mind that the curves
are generated by the differences between two cumulative density functions that have
different parameters. The complex nature of the nonlinearity is illustrated in Figure
11. In this three-dimensional figure, each dark or light band corresponds to .10 on
the vertical axis. Broken lines indicate that the confidence interval at the prediction
crosses 0. If you trace a curve as articles increase from 0 to 50 with prestige held at
5, you obtain the solid curve in Figure 9. Other curves from that figure are created
by holding prestige constant at other levels. The curves in Figure 10 can be created
in a similar fashion.

5 Conclusions
The comparison of groups in regression models for binary outcomes is complicated by
the identification problem inherent in these models. Traditional tests of the equality
of regression coefficients across groups confound the magnitude of the regression coef-
ficients with residual variation. If the amount of variation differs between groups, the
7
Color versions of these graphs are available at www.indiana.edu/~jslsoc/research_groupdif.htm.

11
test can lead to incorrect conclusions. The test proposed by Allison (1999) takes into
account group differences in residual variation by incorporating the assumption that
the effects of some variables are the same across groups. In practice, researchers might
not have a basis for making this assumption, in which case the conclusions from this
test could also be misleading. The approach advocated in this paper is to compare
predicted probabilities across groups as levels of one variable change holding other
variables constant. Since predicted probabilities are not affected by residual varia-
tion, testing the equality of predictions is not affected by the identification problem.
Using predicted probabilities is not, however, without costs. Tests of the equality
of regression coefficients across groups requires a single test and provides a simple
answer to the question of whether a variable has the same effect for both groups.
With predicted probabilities, you must examine group differences at many values to
see how groups differ in the way that changes in predictors affect the outcome.
My proposal for comparing groups allows you to show if there are significant dif-
ferences in the predicted outcomes and whether these differences vary by the level
of the independent variables. This information does not correspond to a test of the
equality of regression coefficients across groups. In models such as logit or probit,
tests of group differences in probabilities do not allow us to determine whether these
differences are due to differences between groups in their regression coefficients, to
differences in residual variation, or to a combination of both. However, using proba-
bilities allows us to show if there are significant differences in the predicted outcomes
and whether these differences vary with the level of the independent variables. For
many purposes, this information is critical for understanding the process being stud-
ied.
Using group differences in predictions requires a great deal of planning about which
comparisons are most interesting and at which values you want to hold other variables
in the model constant. As with logit and probit in general, the effect of a variable (in
this case the effect of group membership) depends on the level of all variables in the
model. Further, as you plan where you want to examine the predictions of your model,
it is important to limit you explorations to those regions of the data space that are
both substantively reasonable and where there are a sufficient observations to justify
your conclusions. To illustrate this problem, consider Model 3 and the graphs showing
gender differences in tenure by departmental prestige and the number of articles.
While these differences are largest for those with high productivity in distinguished
departments, the sample used was for scientists with degrees in the late 1950s and
early 1960s. During this period there were few women in prestigious departments and
hence the region of the data space where we find the largest gender differences has
relatively few individuals. Still, the results are strong and consistent with historical
findings about tenure at distinguished departments during this period (Rossiter 1982,
1995).
Finally, the approach advocated in this paper is not limited to binary logit and

12
probit. The same methods can be used with ordinal and nominal models where similar
issues of identification occur. Moreover, our approach is also valuable in models where
it is possible to test regression coefficients across groups, such as a linear regression
model. Examining group differences in the predicted outcome at specific values may
provide insights that cannot be gained from a single test comparing coefficients.
Initial draft: February 2005
Revised draft: June 2009

6 References
Allison, Paul D. 1999. Comparing Logit and Probit Coefficients Across Groups.
Sociological Methods and Research 28:186-208.
Cameron, A. C. and P. Trivedi. 2005. Microeconometrics: Methods and Applica-
tions. New York, Cambridge.
Chow, G.C. 1960. Tests of equality between sets of coefficients in two linear regres-
sions. Econometrica 28:591-605.
Garfield, Eugene. 1984. This week’s citation classic: “Chow, G.C. 1960. ‘Tests
of equality between sets of coefficients in two linear regressions.’ Econometrica
28:591-605.” Current Contents 49:16.
Liao, Tim Futing. 2002. Statistical Group Comparison. New York: Wiley.
Long, J. Scott. 1997. Regression Models for Categorical and Limited Dependent
Variables. Thousand Oaks, CA: Sage Press.
Long, J. Scott. 2009. GC: Programs for Group Comparisons of Regression Mod-
els for Binary Outcomes. Available running findit scottlong in Stata and
following instructions to download the commands.
Long, J.S. and Freese, J. 2005. Regression Models for Categorical and Limited De-
pendent Variables with Stata. Second Edition. College Station, TX: Stata Press.
Long, J. Scott, Paul D. Allison, and Robert McGinnis. 1993. Rank Advancement in
Academic Careers: Sex Differences and the Effects of Productivity. American
Sociological Review 58:703-722.
Maddala, G.S. 1983. Limited-dependent and Qualitative Variables in Econometrics.
Cambridge: Cambridge University Press.
Rossiter, M.W. 1982. Women Scientists in America: Struggles and Strategies to
1940. Baltimore, Md., Johns Hopkins Press.

13
 Rossiter, M.W. 1995. Women Scientists in America: Before Affirmative Action
1940-1972. Baltimore, Johns Hopkins Press.

StataCorp. 2007. Stata Statistical Software. Release 10. College Station, TX: Stata
Corporation.

Williams, Richard. 2009. Using Heterogeneous Choice Models to Compare Logit

and Probit Coefficients Across Groups. Sociological Methods & Research 37(4):
531-559.

Xu, J. and J.S. Long. 2005. Confidence intervals for predicted outcomes in regression
models for categorical outcomes. The Stata Journal 5: 537-559.

7 Appendix
The test in the paper is a special case of the more general test considered here. The
logit or probit models are defined as

Pr ( = 1 | x) =  (x0 β)

where  () is the cdf for the standard normal distribution for the probit model and the
standardized logistic distribution for the logit model. Let  (β) be the difference in
the predicted probabilities at two sets of values of the independent variables contained
in the vectors x and x :

 (β) = Pr ( = 1 | x ) − Pr ( = 1 | x )
=  (x0 β) −  (x0 β) 

This formulation is more general than the case considered in the text since it can be
applied to any two sets of values of x, not only comparisons between groups. For
example, if the model is

Pr ( = 1 | x) =  ( 0 +  1 1 +  2 2 +  3 3 )

for a single group,  (β) could be the difference in the predicted probability when
1 = 1, 2 = 2, and 3 = 2 and when 1 = 2, 2 = 3, and 3 = 1.
The paper estimates a model in which dummy variables and interactions are used
to include both groups in the same equation. For example, consider a single indepen-
dent variable  and two groups. Let  = 1 for women, else 0; let  = 1 for men, else
0; and let  =  ×  and  =  × . Next, estimate the model

Pr ( = 1 | x) = Λ (   +    +    +   )

14
where no constant is included. For women,  = 1 and  = 0 which simplifies the
equation to
Pr ( = 1 | x) = Λ (  +   )
and for men  = 0 and  = 1 so that
Pr ( = 1 | x) = Λ (  +   ) 
Group difference in the predicted probabilities at a given value of  equal
³ ´
 βb = Pr ( = 1 |  = 1  = 0 ) − Pr ( = 1 |  = 0  = 1 )
³ ´
b b b b b b b b
= Λ(   +    +   0 +   0) − Λ   0 +   0 +    +   
³ ´
= Λ( b + b ) − Λ b + b  

Returning to the general formulation of the problem,

(β)  (x0 β)  (x0 β)
= −
β β β
where
 (x0 β) h Λ(x0 ) Λ(x0 ) Λ(x0 )
i0
=  0  1
···   =  (x0 β) x
β
Then,
∙ ¸
 (β|x ) b  (β|x )
Var [Pr ( = 1 | x ) − Pr ( = 1 | x )] =  (β)
β 0 β
∙ ¸
Λ(β|x ) b Λ(β|x  )
+  (β)
β0 β
∙ ¸
 (β|x ) b  (β|x )
−2  (β) 
β0 β
The last term in the equation drops out when interactions are used to include both
groups in the same equations and x and x are used to specify the values for each
group. A −statistic to test 0 : Pr ( = 1 | x ) = Pr ( = 1 | x ) can be computed
as
Pr ( = 1 | x ) − Pr ( = 1 | x )
=p
Var [Pr ( = 1 | x ) − Pr ( = 1 | x )]
which has an asymptotic normal distribution. Or, we can construct a confidence
interval around ∆ (x  x ) = Pr ( = 1 | x ) − Pr ( = 1 | x ). For example, the 95%
confidence interval for the difference in probabilities ∆ (x  x ) includes upper and
lower bounds such that
Pr (∆ (x  x )LB ≤ ∆ (x  x ) ≤ ∆ (x  x )UB ) = 95

15
 Since the covariance term  
(|x )
0
b  (|x ) is zero when comparing groups,
 (β) 
any software that computes confidence intervals and the standard error of the predic-
tion can be used. For the more general case, the SPost commands (Long and Freese
2005; Xu and Long 2005) can be used for logit and probit as well as other models.

16
8 Tables and Figures

Table 1: Descriptive statistics for male and female biochemists using person-year
observations.

Women Men Combined

Variable Mean SD Mean SD Mean SD
Is tenured? tenure 0.109 0.312 0.132 0.338 0.123 0.328
Year year 3.974 2.380 3.784 2.252 3.856 2.303
Year-squared yearsq 21.457 23.137 19.388 21.501 20.169 22.151
Bachelor’s selectivity select 5.001 1.475 4.992 1.365 4.995 1.407
Total articles articles 7.415 7.430 6.829 5.990 7.050 6.576
Distinguished job? distinguished 0.058 0.233 0.040 0.197 0.047 0.211
Prestige of job prestige 2.658 0.765 2.640 0.784 2.647 0.777
N 1,056 1,741 2,797

Table 2: Logit model 1 predicting tenure for male and female biochemists using only
the number of articles.
Women Men
Variable  exp ()   exp () 
constant -2.501 -17.96 -2.721 -22.4
articles 0.047 1.048 4.49 0.102 1.102 9.76
log-likelihood -353.758 -628.282
N 1,056 1,741

17
Table 3: Logit model 2 predicting tenure for male and female biochemists using
articles and an indicator of having a high prestige job.

Women Men
Variable  exp ()   exp () 
constant -2.604 -17.32 -2.715 -22.27
articles 0.068 1.070 5.36 0.106 1.111 9.89
distinguished -1.984 0.138 -2.69 -0.945 0.389 -2.06
log-likelihood -348.344 -625.672
N 1056 1,741

Table 4: Logit models predicting tenure for male and female biochemists.
Women Men
Variable  exp ()   exp () 
constant -5.842 -6.75 -7.680 -11.27
year 1.408 4.087 5.47 1.909 6.745 8.92
yearsq -0.096 0.909 -4.36 -0.143 0.867 -7.70
select 0.055 1.057 0.77 0.216 1.241 3.51
articles 0.034 1.035 2.69 0.074 1.076 6.37
prestige -0.371 0.690 -2.38 -0.431 0.650 -3.96
log-likelihood -306.191 -526.545
N 1056 1741

18
Figure 1: The link between  ∗ = −6 + 1 +  with  = 1 and Pr ( = 1 | ) for
parameter Set A.

Figure 2: The link between  ∗ = −12 + 2 +  with  = 2 and Pr ( = 1 | ) for

parameter Set B.

19
Figure 3: Changes in probabilities for men and women with the same coefficient for
 but different intercepts.

20
Figure 4: The probability of tenure for men and women by the number of articles
using Model 1.

Figure 5: Gender differences in the probability of tenure by the number of articles

using Model 1.

21
Figure 6: Gender differences in the probability of tenure by the number of articles for
scientists who are not in distinguished departments using Model 2.

Figure 7: Gender differences in the probability of tenure by the number of articles for
scientists who are in distinguished departments using Model 2.

22
Figure 8: Gender differences in the probability of tenure by the number of articles
and by prestige of department using Model 2.

23
Figure 9: Gender differences in the probability of tenure by the number of articles at
different levels of departmental prestige using Model 3.

Figure 10: Gender differences in the probability of tenure by prestige at different

levels of productivity using Model 3.

24
Figure 11: Gender differences in the probability of tenure by prestige and number of
articles using Model 3.

25